Possible divergence of the method krige.conv ?

[pierromond]

Hi,

I am using your very useful code, and thank you for it !

Nevertheless, when I am using the krige.conv method, I have some problem of divergence that I don't observe using the standard krige.conv method of GeoR with the same parameters.

Adding the distance matrices give me very diverging results when the geodata set increase. For one or two points to krige on my network I don't have any problem, but when there are more than 500 points to "merge", very extreme values appear.

Did you already observe this phenomenon ? Do you have any idea about which parameters could be badly given in my function ?

krige.conv(geodata= Obs, locations = PT_gd$coords, krige=krige.control(cov.model="spherical",cov.pars=c(30,500), nugget=0), dd.dists.mat=distances, dl.dists.mat=distances_to_obs )

Thank you.

Kind regards,

Pierre

P.S. I checked my distance matrices and they seems to be ok.

[famuvie]

Hi Pierre, thanks for your comments. It is hard to say without looking at the data. I have not observed the problem you report. Here are some questions/things to check:

• I'm not sure what you mean by "merging" points. Do you mean predicting in more points?

• Why do you speak of a "network"?

• What do you mean by "divergence"? you mean that the prediction algorithm does not converge? or rather that the predictions are too much different from regular euclidean kriging predictions?

• Does your covariance model fits reasonably well your cost-based empirical variogram?

Best wishes ƒacu.-

[pierromond]

You are alright,

I will try to precise my question.

• I'm not sure what you mean by "merging" points. Do you mean predicting in more points? Yes, I do. "predict" is much better than "merge" :)

• Why do you speak of a "network"? Because all my observations and predictions locations are above a network (road network). I have distance matrices calculated between each locations from the graph of the road network.

• What do you mean by "divergence"? you mean that the prediction algorithm does not converge? or rather that the predictions are too much different from regular euclidean kriging predictions? I am working with sound levels. If I use 5 or 10 observations all my predicted points are between 50 and 90 dB and the results seems coherent, but upon 500 measurements, I observe sound levels predicted between -1000 and 1000 dB(A). When I use the euclidian configuration with 500 measurements, I still have coherent results.

• Does your covariance model fits reasonably well your cost-based empirical variogram? Yes the model fits quite well.

Maybe could I share some part of code and data with you ?

Greetings,

Pierre

[famuvie]

Strange, since no matter the number of observations, any prediction ends up being a weighted average of your observations. Are you sure there are no errors in you larger set of measurements?

I can have a look if you want, but I don't guarantee anything, since I am very short of time these days.

[pierromond]

I tried to find errors in my set but I didn't. And because it works with the standard Krige.conv, it lets me think that it come from another source of errors. Maybe distance matrices, but I can't find.

Joined you can find a extract of my data set and script with my comments (Test.R and test.Rdata). If you can find some time, it could be great, I tried to send you the more condensed version of my script! By the way, I will continue to check my code, and read again your publication to find my mistake.

R.zip

Thank you again,

Pierre Best wishes,

Pierre

[pierromond]

Hi,

Finally, I used exactly the same data format than you and now it works.

Thank you a lot for your work ! and I will close the issue.

Best wishes,

Pierre

[famuvie]

Ops, you were faster than me :) What do you mean by data format? What was the problem then?

[pierromond]

I don't really understand where was my mistake, because I am using exactly the same datas. But for example, I was not using the function coordinates() to give my locations, and maybe my data frame was not the good one, or sthg like that...